14 DAY TRIAL // Google acquired DeepMind earlier this year to help with its Artificial Intelligence efforts. After revealing efforts to expand the team by hiring some top-notch scientists, DeepMind has unveiled that it has built a neural network that can access an external memory, like the Turing machine.
Basically, the result of this new endeavor is a computer that mimics the short-term memory of the human brain, which is one big step towards the final goal of creating Artificial Intelligence.
The researchers at DeepMind have been working on the so-called Neural Turing Machine, which combines a neural network controller with a memory bank, which gives it the ability to learn to store and retrieve information. The name, as you may have picked up already, refers to Alan Turing, a computer pioneer who imagined computers as machines that have a working memory for storage and retrieval of data.
What DeepMind has managed to do is a breakthrough after a long history of work on this exact type of feature that started decades ago. While theories and neural networks have been around for quite some time, there’s renewed attention for them as more powerful computers seek to take advantage of them.
It’s not just Google that’s after this gold mine, but also Facebook, which recently announced that it had managed to train a neural network to identify faces with near-human accuracy.
Unmatched achievements
At DeepMind, researchers put the Neural Turing Machine through a series of tests, which included tasks such as copying and sorting blocks of data. Compared to a conventional neural net, DeepMind managed to build one that learns faster and copies longer data sequences with fewer errors.
“Our experiments demonstrate that it is capable of learning simple algorithms from example data and of using these algorithms to generalize well outside its training regime,” write DeepMind’s Alex Graves, Greg Wayne, and Ivo Danihelka in a research paper.
This is particularly significant because the conventional neural network learns to copy sequences up to length 20 with near perfection. Longer sequences that exceed what the network has been trained to deal with return errors. The Neural Turing Machine managed to copy a sequence of length 120.
Google acquired DeepMind earlier this year for $400 million (€317 million). The company is based in London and was founded in 2011 as an artificial intelligence company. The purchase came after Google had already invested quite a lot of money into several robotics companies.